# Reverse voice anonomizing algorithMs

In the movie saw, the voice of the recordimg tapes was a process similiar to having 2 identical recorded speach track which ones pitch is changed down some percent and the other tracks pitch is lowered more and then combined together. For instamce

-20% track a -30% track b

I tried to undo this with software by upping the pitch by the average of the 2 pitch deductions 25% but it doesnt sound high enough. How do you reverse such an algorithm?

-

## migrated from avp.stackexchange.comJan 27 '14 at 15:10

This question came from our site for engineers, producers, editors, and enthusiasts spanning the fields of video, and media creation.

out of curiosity, who are you trying to expose? I saw your post here avp.stackexchange.com/q/7334/3592. – Jason Conrad Mar 12 '13 at 3:42
Just saw special effect and how they reversed them and was curious how @Jason Conrad – Chris Okyen Mar 12 '13 at 14:48
The special effects of reversing the saw voice of jogsaw got me curious – Chris Okyen Mar 12 '13 at 17:12

Basically, if you mix two different frequencies (and this is the outcome of mixing audio with different pitch), you will get a multitude of frequencies.

For example, if the original audio contains a frequency of 1000 hz at some point, this will be represented as 800 Hz and 700 Hz. If you now mix these frequencies, you will get the sum of these frequencies (1500 Hz) as well as the difference (100 Hz) in the resulting output. Now with this multitude of frequencies (100, 700, 800 and 1500 Hz) which are all on different levels, the signal is really distorted and the source signal is hard to reconstruct (probably what all those evildoers will want).

As long as you have a single frequency you want to reconstruct, you might use up- and downmixing and then filtering out the leftover parts of the signal to regain the original frequency. This is what a radio usually does to get the audio signal from the radio frequency.

If you have a mixture of frequencies that are changing all the time, it is nearly impossible to reconstruct the original audio, as you can only guess and try and filter, as you already did.

-
"you will get the sum of these frequencies (1500 Hz) as well as the difference (100 Hz) in the resulting output" only if you pull the signal through something nonlinear; as long as you just play the clean mixture you only get the frequencies you put in and these are seperable without much ado. What is a problem is that speech is not just "contains a frequency of 1000 Hz" but is itself quite a wild mixture. Nevertheless... – leftaroundabout Mar 2 '13 at 12:23

The following assumes the effect really works as you think. I don't actually know!

Before you start undoing any pitch-shift you will need to separate the tracks. This can't be done properly with ordinary EQing, but should be rather accomplishable with a comb filter, since speech (at least the parts that respond critically to a pitch-shift) consist mostly of "ordinary" harmonic overtones, i.e. frequency-multiples of some common fundamental. Unfortunaly, this fundamental frequency is not quite constant, it needs to be tracked dynamically. You might be able to employ a tool like Melodyne, but it only works for sure with manual trial-and-error-based tracking – an autocorellation or at least FFT spectrum should make this a lot easier – or specialised tools like RX or reNOVAtor.

Once you have seperated the channels, the rest should be easy.

-
muffling the voice as they did in the saw franchise does work if thats what you mean by "assumig the effect works." Did you mean if undoing this works? – Chris Okyen Mar 4 '13 at 19:51
I meant, if the anonymisation has the origin as you ascribed (multiple channels of different pitch-shifting), rather than some other process. – leftaroundabout Mar 4 '13 at 22:19